A deep learning enhanced protein search engine optimization method and system
By employing a deep learning feature fusion mechanism that combines hash indexing, cached batch processing, and dual-decoy calibration, the protein search engine overcomes the identification bottleneck in high-complexity data scenarios, achieving high-precision, high-sensitivity, and reliable protein identification, applicable to various search paradigms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF COMPUTING TECH CHINESE ACAD OF SCI
- Filing Date
- 2026-05-19
- Publication Date
- 2026-07-31
AI Technical Summary
Existing protein search engines suffer from systemic technical defects in high-complexity data scenarios, including insufficient peptide matching accuracy, limited discrimination capabilities of traditional rescoring frameworks, and the inability of deep learning rescoring to adapt to open search scenarios, making it difficult to balance identification sensitivity and reliability.
A high-precision coarse scoring module employing hash indexing and cached batch processing, combined with the XGBoost rescoring framework with dual decoy calibration and a trust-aware deep learning feature fusion mechanism, achieves accurate matching between theoretical fragment ions and experimental spectral peaks and deep learning-enhanced rescoring, supporting a unified solution for both restricted and open search paradigms.
It significantly improves the accuracy and sensitivity of candidate peptide screening while maintaining strict control over the false discovery rate, thereby enhancing the reliability and sensitivity of protein identification, especially in high-density spectra and complex sample analysis.
Smart Images
Figure CN122493945A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of bioinformatics and computational proteomics, specifically to a deep learning-enhanced protein search engine optimization method and system. Background Technology
[0002] Proteomics aims to systematically characterize proteins in biological samples, including their identity, expression levels, and post-translational modifications (PTMs), thereby revealing the molecular mechanisms of life activities and providing a basis for the discovery of disease biomarkers. Liquid chromatography-tandem mass spectrometry (LC-MS / MS) technology has become the cornerstone of modern proteomics due to its high resolution and high accuracy. In a typical workflow, proteins are enzymatically digested into peptides, separated by liquid chromatography, and then enter a mass spectrometer, where they are ionized and further fragmented to produce MS / MS spectra. Protein search engines, as the core computational tool for translating mass spectrometry data into biological information, directly determine the depth and reliability of analysis.
[0003] Database search is currently the most important and widely used computational method for proteomics identification. This method compares the similarity of experimental MS / MS spectra with virtual peptide spectra generated from theoretical enzyme digestion based on protein sequence databases. A scoring function is used to assess the matching quality, identifying the most likely peptide identity and thus achieving protein identification. Based on different methods of handling post-translational modifications, it is mainly divided into the following two categories:
[0004] Limited Search: The core of this method is to pre-limit the search space to ensure computational efficiency. Its basic process is as follows: First, theoretical enzyme digestion is performed on the protein sequence database according to enzyme digestion rules. Possible modifications are enumerated on the peptide sequences, and candidate fragments are quickly matched to the spectra using the parent ion mass tolerance. Then, based on mass spectrometry fragmentation mechanisms (such as HCD primarily producing b / y type ions), the mass-to-charge ratio of theoretical fragment ions is calculated for each candidate peptide. The experimental spectra are compared with the theoretical spectra using a specialized scoring function (such as algorithms based on cross-correlation or probabilistic models) to evaluate the mass matching accuracy, intensity correlation, and sequence coverage continuity of the fragment ions, achieving efficient identification.
[0005] Open search: These methods generally employ candidate screening strategies based on sequence tags or ion indices to achieve efficient traversal of ultra-large-scale search spaces. By extracting short sequence tags or establishing ion indices from experimental MS / MS spectra, the range of candidate peptides is narrowed from the complete sequence plus modification space to a small subset containing specific sequence patterns or quality patterns, thus avoiding the computational explosion caused by exhaustively searching all possible modification combinations. Building on this, open search methods establish correspondences between candidate peptides and experimental spectra using sequence-spectrum alignment algorithms that tolerate quality shifts (such as dynamic programming or index-based matching), allowing for significant differences between peptide sequences and parent ion qualities (i.e., potential modification quality), and evaluating modification positions and matching quality through scoring functions.
[0006] Subsequently, a re-scoring strategy is typically employed to further enhance identification confidence and sensitivity. The core principle is to transform the peptide-spectrum matching (PSM) classification problem into a statistical learning problem based on multi-dimensional features, usually following a "feature engineering - semi-supervised training - re-ranking" technical route. First, descriptive features, such as the search engine's raw score and quality error, are extracted from the output of the search engine's pre-processing stage. Then, using a semi-supervised learning framework, a binary classifier is trained to learn the feature distribution boundaries of correct and incorrect matches, with high-confidence target PSMs as potential positive samples and decoy PSMs as explicit negative samples, thus re-ranking the candidate PSMs. In this stage, a data-driven re-scoring method further increases the number of identifications by introducing deep learning features. It uses neural networks to predict verifiable attributes such as peptide retention time and fragment ion intensity, using the deviation between experimental observations and predicted values as consistency features. This provides strong discriminative features independent of traditional scoring functions, thereby improving the model's discriminative ability.
[0007] Finally, statistical validation is often performed using the target-bait library method (TDA) to estimate the false discovery rate (FDR) of peptide identification in order to output high-quality identification results.
[0008] The problems and shortcomings of existing technologies include:
[0009] In the peptide spectrum matching and scoring stage, search engines such as pFind and Comet generally adopt a vectorization strategy: by discretizing the continuous mass-to-charge ratio (m / z) axis into several bins, the signal intensities of experimental spectra and theoretical fragment ions are aggregated into the corresponding bins, and then the degree of peptide spectrum matching is quantified by vector similarity. This method is computationally simple and efficient, but its coarse-grained binning mechanism is prone to introducing mismatches; especially in scenarios with dense spectral peaks generated by high-precision instruments, the negative impact on identification sensitivity is more significant. In contrast, ion indexing technology, represented by MSFragger, achieves orders of magnitude improvement in search speed, but comes with huge storage overhead (e.g., in open searches, the index of HLA peptides or phosphorylation modifications can reach tens of gigabytes), and this method lacks the ability to detect fragment ions generated by non-predefined modifications, thus limiting the sensitivity of modification identification.
[0010] In the rescoring phase, traditional frameworks such as Percolator primarily employ Support Vector Machines (SVMs) as classifiers. However, real-world PSM data often contains complex nonlinear feature interactions, and the decision boundaries of linear SVMs are insufficient to fully model such relationships, typically relying on explicit kernel function engineering to extend their expressive power. Although some methods have begun to adopt more expressive learning models like XGBoost as classifiers to enhance nonlinear fitting capabilities, simply migrating the original training methods to the new classifiers, with their stronger fitting capabilities, is more prone to false positive underestimation problems.
[0011] For data-driven rescoring, existing tools such as MSBooster, Oktoberfest, and MS2Rescore are only designed for restricted searches. This is because their peptide attribute prediction models only support a limited number of predefined modification types and are not supported in open search phases. Although DeepLC uses atomic composition encoding to predict the retention time of peptides with unknown modifications in open searches, it is currently only used as an auxiliary verification method to screen and identify results and has not yet been systematically integrated into the rescoring framework to support rescoring under open searches, thus limiting the ability to discover unexpected modifications.
[0012] In summary, the current field still lacks a unified search framework that is compatible with both restricted and open search paradigms and systematically supports deep learning-enhanced rescoring under both modes. Summary of the Invention
[0013] The purpose of this invention is to overcome the systemic technical defects of existing protein search engines in high-complexity data scenarios: First, existing peptide spectrum matching and scoring algorithms sacrifice accuracy for speed, making it difficult to accurately screen candidate peptides from high-density spectra, thus becoming a bottleneck restricting identification sensitivity; second, the linear models used in traditional rescoring frameworks have limited discriminative power, while directly introducing high-performance nonlinear models can lead to the failure of FDR control, creating an inherent contradiction between "sensitivity" and "reliability"; third, deep learning rescoring cannot adapt to open search scenarios. To solve the above problems, this invention proposes a deep learning-enhanced protein search engine multi-module collaborative optimization method and system, which can significantly enhance the sensitivity of the protein search engine while maintaining reliable error rate control, and provides a unified solution for deep coverage and discovery of unknown modifications in high-complexity mass spectrometry data.
[0014] In a first aspect, embodiments of this application provide a deep learning-enhanced protein search engine optimization method, the method comprising:
[0015] Peptide spectrum matching coarse scoring steps: For the mass spectrometry data of proteins, hash index is used in peptide spectrum matching to perform accurate matching calculations between theoretical fragment ions and experimental peaks within a set tolerance. A batch processing peptide caching mechanism and a dedicated memory pool management strategy are used in the calculation.
[0016] Peptide profile matching rescoring step: For the mass spectrometry data that has been coarsely scored, a dual-decoy calibration strategy is adopted to divide the decoys into a mutually exclusive visible queue for training and a hidden queue for false discovery rate estimation. A calibration factor is introduced to correct the false discovery rate estimation and output the rescored data.
[0017] Deep learning steps: Based on the rescoring data, the physicochemical properties of peptides are predicted using a deep learning model. Consistency features are constructed by comparing with experimental observations, and a trust-aware mechanism is adopted to achieve deep learning-enhanced rescoring in open scenarios.
[0018] In this embodiment of the invention, the above-mentioned peptide profile matching coarse scoring step includes:
[0019] Candidate enumeration and matching steps: The protein database is theoretically digested, and possible modification combinations are enumerated within a specified modification list to generate theoretical peptides. Based on the parent ion mass tolerance, the theoretical peptides are rapidly matched and correlated with experimental spectra.
[0020] Batch scoring steps: Whenever a peptide matches a spectrum, it is directly added to the cached peptide list of that spectrum. When the cached list of a spectrum reaches a predetermined capacity, a batch coarse scoring is triggered, and the peptides are added to the candidate peptide list of that spectrum in turn, while the cached peptide list is cleared.
[0021] Memory management steps: A memory management mechanism is adopted to initialize a user-mode memory pool for each thread. Its size is dynamically expanded at runtime, and it supports the creation and destruction of peptide objects.
[0022] In this embodiment of the invention, the above-mentioned peptide profile matching and rescoring step includes:
[0023] Dual-decoy segmentation steps: Based on open and restricted search results, initialize all bait peptide-spectrum matching PSMs and divide them into hidden baits and visible baits of equal size and mutual exclusion according to certain rules.
[0024] Rescoring iteration steps: Select non-visible decoys within a specified false discovery rate threshold as positive samples, filter some visible decoys as negative samples, integrate their features, and train the rescoring model.
[0025] The trained model is used to predict all samples, and the samples are reordered based on the prediction results. The false discovery rate is estimated using hidden decoys. This process is repeated multiple times until convergence or a specified number of times is reached.
[0026] In this embodiment of the invention, the above-mentioned deep learning steps include:
[0027] The pre-trained model was fine-tuned using a small number of high-confidence samples to predict the retention time and fragment ion intensity of candidate peptides.
[0028] Alternatively, an acceleration strategy or model quantization and half-precision inference techniques can be used to calculate the consistency characteristics between predicted values and experimental observations.
[0029] For open search scenarios, unconventional modifications outside the scope of data coverage are fine-tuned by using additional features as indicators of reduced weighting, automatically lowering their prediction confidence weights.
[0030] Secondly, embodiments of this application provide a deep learning-enhanced protein search engine optimization system, employing the deep learning-enhanced protein search engine optimization method described above. The system includes:
[0031] Peptide spectrum matching coarse scoring module: For protein mass spectrometry data, the peptide spectrum matching module uses hash index to perform accurate matching calculations between theoretical fragment ions and experimental peaks within a set tolerance, and employs a batch processing peptide caching mechanism and a dedicated memory pool management strategy in the calculation.
[0032] Peptide spectrum matching rescoring module: For mass spectrometry data that has been coarsely scored, a dual-decoy calibration strategy is adopted, which divides the decoys into a mutually exclusive visible queue for training and a hidden queue for false discovery rate estimation, and introduces a calibration factor to correct the false discovery rate estimation, and outputs rescored data.
[0033] Deep learning module: Based on rescoring data, it uses a peptide attribute prediction model to predict features, constructs consistent features by comparing them with experimental observations, and adopts a trust-aware recognition mechanism to achieve deep learning-enhanced rescoring in open scenarios.
[0034] In this embodiment of the invention, the above-mentioned peptide spectrum matching coarse scoring module includes:
[0035] Candidate Enumeration and Matching Module: Performs theoretical enzymatic digestion on the protein database, enumerates possible modification combinations in the specified modification list, generates theoretical peptides, and rapidly matches and associates theoretical peptides with experimental spectra based on the parent ion mass tolerance.
[0036] Batch scoring module: Whenever a peptide matches a spectrum, the peptide is added to the spectrum's cached peptide list. When the cached list of a spectrum reaches a predetermined capacity, a batch coarse scoring is performed, and the peptides are added to the spectrum's candidate peptide list in turn, while the cached peptide list is cleared; thus realizing that each spectrum corresponds to a batch of peptides.
[0037] Memory management module: Employs a memory management mechanism to initialize a user-mode memory pool for each thread. The size of the memory pool expands dynamically at runtime and supports the creation and destruction of peptide objects.
[0038] In this embodiment of the invention, the peptide profile matching rescoring module includes:
[0039] Dual-decoy segmentation module: Based on open and restricted search results, initialize all decoy peptide-spectrum matching PSMs and divide them into hidden decoys and visible decoys of equal size and mutual exclusion according to certain rules;
[0040] Rescoring Iteration Module: Selects non-visible decoys within a specified false discovery rate threshold as positive samples, filters some visible decoys as negative samples, integrates their features, and trains the rescoring model.
[0041] The trained model is used to predict all samples, and the samples are reordered based on the prediction results. The false discovery rate is estimated using hidden decoys. This process is repeated multiple times until convergence or a specified number of times is reached.
[0042] In this embodiment of the invention, the deep learning module includes:
[0043] The pre-trained model was fine-tuned using a small number of high-confidence samples to predict the retention time and fragment ion intensity of candidate peptides.
[0044] Alternatively, an acceleration strategy or model quantization and half-precision inference techniques can be used to calculate the consistency characteristics between predicted values and experimental observations.
[0045] For open search scenarios, unconventional modifications outside the scope of data coverage are fine-tuned by using additional features as indicators of reduced weighting, automatically lowering their prediction confidence weights.
[0046] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a deep learning-enhanced protein search engine optimization method.
[0047] Fourthly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the deep learning-enhanced protein search engine optimization method described above.
[0048] Compared with existing technologies, it has the following outstanding advantages:
[0049] 1) The method and system of this invention propose a high-precision coarse scoring module based on hash indexing and cached batch processing. This module abandons the traditional vectorized approximate matching and uses hash indexing to achieve accurate matching between theoretical fragment ions and experimental spectral peaks within a set tolerance. It also designs a cached batch processing and dedicated memory pool management strategy to offset the computational overhead brought about by the improvement in accuracy.
[0050] Technical results: While maintaining computational efficiency similar to the baseline method, the accuracy of candidate peptide screening was significantly improved, providing a higher quality and more reliable candidate set for the downstream rescoring module.
[0051] 2) The method and system of this invention propose an XGBoost rescoring framework based on dual-decoy calibration. This framework uses an XGBoost ensemble learning model to replace the traditional linear classifier, in order to fully model the complex feature interactions in PSM. To address the statistical bias problem caused by this, a dual-decoy calibration strategy is designed, dividing the decoy spectral matching into two queues: "visible" and "hidden," which are used for model training and unbiased error detection rate estimation, respectively.
[0052] Technical effects: While unleashing the powerful discriminative potential of nonlinear models and significantly improving sensitivity, it ensures the strict unbiasedness of the false detection rate estimation and solves the imbalance between high-discrimination models and statistical reliability.
[0053] 3) The method and system of this invention propose a trust-aware deep learning feature fusion mechanism for open search. This mechanism systematically integrates the pPred deep learning model, which supports predictions of over a thousand modifications, into the open search process. It constructs a consistency metric between predicted features and experimental observations, and introduces trust-aware identification to deweight predicted outputs with rare or low-confidence modifications.
[0054] Technical results: For the first time, a stable and reliable application of deep learning enhancement in open search scenarios has been achieved, further improving the identification sensitivity in open search scenarios. Attached Figure Description
[0055] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0056] Figure 1 This is a schematic diagram of the deep learning-enhanced protein search engine optimization method of the present invention.
[0057] Figure 2 This is a schematic diagram of a deep learning-enhanced protein search engine optimization method according to an embodiment of the present invention;
[0058] Figure 3 This is a schematic diagram of the deep learning-enhanced protein search engine optimization system of the present invention.
[0059] Figure 4 This is a schematic diagram of the computer hardware of the present invention. Detailed Implementation
[0060] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.
[0061] It should also be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.
[0062] It should also be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0063] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.
[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0065] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0066] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0067] To make the above-mentioned features and effects of the present invention clearer and easier to understand, specific embodiments are described below in conjunction with the accompanying drawings. This specification discloses one or more embodiments incorporating the features of the present invention. The disclosed embodiments are merely illustrative. The scope of protection of the present invention is not limited to the disclosed embodiments, but is defined by the appended claims.
[0068] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.
[0069] While conducting in-depth proteomics data analysis research, the inventors discovered that existing search engines face significant challenges in processing data generated by next-generation high-sensitivity mass spectrometers (such as the Orbitrap Astral). This data exhibits extremely high peak density and a wider dynamic range, making traditional algorithms insufficiently adaptable. Furthermore, when analyzing complex samples, the inventors found that existing tools suffer from insufficient sensitivity: the high polymorphism of HLA peptides and non-specific enzymatic cleavage lead to a dramatic expansion of the search space; the multi-species mixing of metaproteomics causes the database size to explode; and the extremely low sample loading at the single-cell level places stringent demands on the peptide profile matching and discrimination capabilities.
[0070] Through analysis of the pFind search engine's stage-by-stage results, the inventors discovered a severe bottleneck in the first stage (coarse scoring) of peptide spectrum matching: the vectorized binning strategy has extremely coarse granularity, leading to serious mismatches in high-density peak scenarios and exacerbating the positional encroachment of high-scoring false positive peptides in large database scenarios, causing a large number of real peptides to be mistakenly screened out in the initial filtering stage; if high-precision peak-by-peak matching is directly adopted, it will trigger random memory access that is not cache-friendly, resulting in a precipitous drop in computational throughput. To address this structural contradiction between "precision and efficiency," the inventors proposed replacing vectorized binning with precise matching based on hash indexes. At the same time, they designed a batch processing peptide caching mechanism to reconstruct the computation and memory access mode, transforming peptide-driven random access into spectrum-driven sequential access, and introducing memory pool management to alleviate the overhead of dynamic memory allocation, thereby restoring high-precision matching capabilities within an acceptable computational cost.
[0071] The significant improvement in upstream candidate peptide screening capabilities places higher demands on the performance of downstream processes. The inventors discovered that the original rescoring framework only considered the statistical interpretability of linear models, neglecting the discriminative power requirements arising from the complex nonlinear interactions between PSM features. While introducing nonlinear models can enhance fitting ability, the inventors found through experiments that this simple transfer can disrupt the statistical commutativity of the target-decoy pair, leading to a systematic underestimation of FDR. To address this issue, the inventors designed a dual-decoy calibration strategy, dividing the decoys into two mutually exclusive queues: visible (for training) and hidden (for FDR estimation). A calibration factor is introduced to correct the FDR estimation, ensuring the unbiasedness of statistical calibration while unleashing the discriminative potential of the strong classifier.
[0072] Considering that deep learning features can further enhance re-scoring discrimination capabilities, the inventors utilized the peptide attribute prediction model pPred to predict the retention time (RT) and fragment ion intensity (MS2) of peptides. By comparing these with experimental observations, consistency features were constructed to capture the physicochemical correlation between spectra and peptide sequences. Simultaneously, the inventors discovered that while existing data-driven tools can handle deep learning re-scoring in restricted search scenarios, they cannot directly adapt to open search scenarios. Therefore, the inventors designed a trust-aware mechanism that identifies rare modifications outside the coverage of fine-tuned data and automatically reduces their prediction confidence weight. This is the first time that deep learning-enhanced re-scoring in open scenarios has been achieved within a unified framework, thereby improving the sensitivity of modification detection.
[0073] The system of this application embodiment will be described in detail below with reference to specific embodiments:
[0074] like Figure 1 and Figure 2 As shown, this invention provides a multi-module collaborative optimization method for a protein search engine enhanced by deep learning. (As per this application...) Figure 1 and Figure 2 As shown, this invention improves the sensitivity of a protein search engine by modifying several key modules within the search engine. The implementation details of these key modules are explained below in conjunction with the overall execution steps.
[0075] like Figure 1 As shown in the embodiment of this application, a deep learning-enhanced protein search engine optimization method is provided, the method comprising:
[0076] Step 101 of peptide spectrum matching coarse scoring: For the mass spectrometry data of the protein, the hash index is used in peptide spectrum matching to perform accurate matching calculation between theoretical fragment ions and experimental peaks within a set tolerance. In the calculation, a batch peptide caching mechanism and a dedicated memory pool management strategy are adopted.
[0077] Peptide profile matching rescoring step 102: For the mass spectrometry data that has been coarsely scored, a dual-decoy calibration strategy is adopted to divide the decoys into a mutually exclusive visible queue for training and a hidden queue for false discovery rate estimation. A calibration factor is introduced to correct the false discovery rate estimation and output the rescored data.
[0078] Deep learning step 103: Based on the rescoring data, use a deep learning model to predict the physicochemical properties of peptides, construct consistency features by comparing with experimental observations, and adopt a trust-aware mechanism to achieve deep learning-enhanced rescoring in open scenarios.
[0079] In this embodiment of the invention, the above-mentioned peptide profile matching coarse scoring step 101 includes:
[0080] Candidate enumeration and matching steps: The protein database is theoretically digested, and possible modification combinations are enumerated within a specified modification list to generate theoretical peptides. Based on the parent ion mass tolerance, the theoretical peptides are rapidly matched and correlated with experimental spectra.
[0081] Batch scoring steps: Whenever a peptide matches a spectrum, it is directly added to the cached peptide list of that spectrum. When the cached list of a spectrum reaches a predetermined capacity, a batch coarse scoring is triggered, and the peptides are added to the candidate peptide list of that spectrum in turn, while the cached peptide list is cleared.
[0082] Memory management steps: A memory management mechanism is adopted to initialize a user-mode memory pool for each thread. Its size is dynamically expanded at runtime, and it supports the creation and destruction of peptide objects.
[0083] In this embodiment of the invention, the above-mentioned peptide profile matching and re-scoring step 102 includes:
[0084] Dual-decoy segmentation steps: Based on open and restricted search results, initialize all bait peptide-spectrum matching PSMs and divide them into hidden baits and visible baits of equal size and mutual exclusion according to certain rules.
[0085] Rescoring iteration steps: Select non-visible decoys within a specified false discovery rate threshold as positive samples, filter some visible decoys as negative samples, integrate their features, and train the rescoring model; use the trained model to predict all samples, re-sort them according to the prediction results, and estimate the false discovery rate using hidden decoys, repeating the process multiple times until convergence or the specified number of times is reached.
[0086] In this embodiment of the invention, the deep learning step 103 includes:
[0087] The pre-trained model was fine-tuned using a small number of high-confidence samples to predict the retention time and fragment ion intensity of candidate peptides.
[0088] Alternatively, an acceleration strategy or model quantization and half-precision inference techniques can be used to calculate the consistency characteristics between predicted values and experimental observations.
[0089] For open search scenarios, unconventional modifications outside the scope of data coverage are fine-tuned by using additional features as indicators of reduced weighting, automatically lowering their prediction confidence weights.
[0090] More specifically, such as Figure 2 As shown in the specific embodiment of this application, a deep learning-enhanced protein search engine optimization method is provided, the method comprising:
[0091] Step S1: If it is an open search, the spectrum is preprocessed, sequence tags are extracted, and peptides in the protein database are matched by extending the sequence, supporting arbitrary modification matching; if it is a restricted search, proceed directly to step S3.
[0092] Step S2: Re-scoring the preliminary results from Step S1 and reducing the protein database based on these results. Simultaneously, statistically analyzing and learning systematic parameters present in the current dataset, such as high-abundance modifications and quality error distribution, as prior knowledge for subsequent restricted searches.
[0093] Step S3: This step is one of the core optimization modules, which aims to achieve both high efficiency and high accuracy in candidate peptide screening.
[0094] Sub-step S31 (Efficient Memory Management): The system employs an efficient memory management mechanism to support the creation and release of massive, high-frequency peptide objects with variable-length parameters. A private memory pool is initialized for each thread, which can be dynamically expanded at runtime, thus avoiding the overhead of frequently requesting and releasing memory from the operating system.
[0095] Sub-step S32 (candidate enumeration and matching): The protein database is theoretically digested, and possible modification combinations are enumerated within a specified modification list to generate theoretical peptides. Based on the parent ion mass tolerance, the theoretical peptides are rapidly matched and correlated with experimental spectra.
[0096] Sub-step S33 (Cached Batch Scoring): Whenever a peptide matches a spectrum, it is directly added to the cached peptide list of that spectrum (the dynamic space required for each peptide object is allocated by the memory pool). When the cached list of a spectrum reaches a predetermined capacity, a batch coarse scoring is triggered (this process uses vector hash indexing to accelerate matching), and the peptides are sequentially added to the candidate peptide list of that spectrum (always retaining a certain number of peptides with the highest scores). At the same time, the cached peptide list is cleared (the dynamic space of the peptide objects reclaimed at this time is returned to the memory pool); if the cached list is not full, the program continues to execute. This design changes the calculation order from "per peptide - per spectrum" to "per spectrum - batch peptides", optimizes memory access locality, and controls calculation latency while ensuring scoring accuracy.
[0097] Step S4: This step is another core optimization module, responsible for uniformly and reliably rescoring the restricted search results generated in step S3 (and the open search results generated in step S2).
[0098] Sub-step S41 (decoy segmentation): Read open and restricted search results, initialize all the decoy PSMs, and segment them into hidden and visible decoys of equal size and mutual exclusion according to certain rules.
[0099] Sub-step S42 (Deep Learning Feature Enhancement, Optional): If deep learning mode is enabled, deep learning features need to be assigned to each sample PSM: the pre-trained model is fine-tuned using a small number of high-confidence samples to predict the retention time and fragment ion intensity of candidate peptides. Optionally, acceleration strategies can be used to improve inference efficiency, such as predicting only the top k candidates for each spectrum, or using techniques such as model quantization (INT8) and half-precision (FP16) inference. Subsequently, consistency features between the predicted and experimental observations are calculated, such as the absolute deviation of retention time and the cosine similarity of fragment ion intensity vectors. In particular, for unconventional modifications outside the coverage of the fine-tuned data in open search scenarios, additional features are used as deweighting indicators for subsequent rescoring models.
[0100] Sub-step S43 (Rescoring Iteration and FDR Calibration): Select non-visible decoy PSMs within the specified FDR threshold as positive samples, filter some visible decoy PSMs as negative samples, integrate their features, and train the XGBoost model; use the trained model to predict all sample PSMs, reorder them according to the prediction results, and estimate the FDR using hidden decoys. Repeat this step multiple times until convergence or the specified number of iterations is reached.
[0101] Step S5: Based on the final rescoring results generated in Step S4, protein inference is performed, and peptides are grouped into proteins. Subsequently, FDR calculation and quality control are performed, ultimately outputting a complete identification result report including peptide sequence, modifications, protein assignment, confidence score, etc.
[0102] As described above, the system and method of the present invention can be implemented well.
[0103] Compared with the prior art, the present invention has the following outstanding advantages and beneficial effects:
[0104] 1. Improved identification sensitivity: Tests on multiple datasets show that, while strictly maintaining a 1% peptide FDR level, this invention improves the total number of peptides identified compared to pFind, and consistently outperforms pFind, MaxQuant, Comet, and MSFragger search engines across various search modes. Specifically:
[0105] 1) In the restricted search, the present invention can identify 12.7-29.3% more peptides than other best-performing search modes.
[0106] 2) In open search, the number of peptides identified by this invention is 8.0-38.4% more than other best-performing search modes.
[0107] 2. While improving performance, computational and statistical costs were successfully controlled:
[0108] 1) Computational efficiency: Although the high-precision matching algorithm itself is more time-consuming, optimized caching and memory management strategies keep the additional computational overhead at a low level. While significantly improving sensitivity, it maintains practical computational efficiency.
[0109] 2) Statistical reliability: The use of the decoy library method for verification proves that the FDR is controlled within a reliable range, ensuring the high credibility of the identification results.
[0110] Compared with existing technologies, the collaborative optimization method proposed in this invention is the first to construct a unified technical framework that simultaneously supports both restricted and open search paradigms and can seamlessly integrate deep learning-enhanced rescoring modules under these two paradigms, providing a complete and universal system-level solution for high-sensitivity and high-reliability analysis of protein mass spectrometry data.
[0111] Secondly, such as Figure 3 As shown, this application embodiment provides a deep learning-enhanced protein search engine optimization system, employing the deep learning-enhanced protein search engine optimization method described above. The system includes:
[0112] Peptide spectrum matching coarse scoring module 201: For protein mass spectrometry data, the peptide spectrum matching uses hash index to perform accurate matching calculations between theoretical fragment ions and experimental peaks within a set tolerance, and employs a batch processing peptide caching mechanism and a dedicated memory pool management strategy in the calculation.
[0113] Peptide spectrum matching rescoring module 202: For the mass spectrometry data that has been coarsely scored, a dual-decoy calibration strategy is adopted to divide the decoys into a mutually exclusive visible queue for training and a hidden queue for false discovery rate estimation. A calibration factor is introduced to correct the false discovery rate estimation and output the rescoring data.
[0114] Deep Learning Module 203: Based on rescoring data, it uses a deep learning model to predict the physicochemical properties of peptides, constructs consistency features by comparing with experimental observations, and adopts a trust-aware mechanism to achieve deep learning-enhanced rescoring in open scenarios.
[0115] In this embodiment of the invention, the above-mentioned peptide profile matching coarse scoring module 201 includes:
[0116] Candidate Enumeration and Matching Module: Performs theoretical enzymatic digestion on the protein database, enumerates possible modification combinations in the specified modification list, generates theoretical peptides, and rapidly matches and associates theoretical peptides with experimental spectra based on the parent ion mass tolerance.
[0117] Batch scoring module: Whenever a peptide matches a spectrum, the peptide is added to the spectrum's cached peptide list. When the cached list of a spectrum reaches a predetermined capacity, a batch coarse scoring is performed, and the peptides are added to the spectrum's candidate peptide list in turn, while the cached peptide list is cleared; thus realizing that each spectrum corresponds to a batch of peptides.
[0118] Memory management module: Employs a memory management mechanism to initialize a user-mode memory pool for each thread. The size of the memory pool expands dynamically at runtime and supports the creation and destruction of peptide objects.
[0119] In this embodiment of the invention, the peptide profile matching rescoring module 202 includes:
[0120] Dual-decoy segmentation module: Based on open and restricted search results, initialize all decoy peptide-spectrum matching PSMs and divide them into hidden decoys and visible decoys of equal size and mutual exclusion according to certain rules;
[0121] Rescoring Iteration Module: Selects non-visible decoys within a specified false discovery rate threshold as positive samples, filters some visible decoys as negative samples, integrates their features, and trains the rescoring model.
[0122] The trained model is used to predict all samples, and the samples are reordered based on the prediction results. The false discovery rate is estimated using hidden decoys. This process is repeated multiple times until convergence or a specified number of times is reached.
[0123] In this embodiment of the invention, the deep learning module 203 includes:
[0124] The pre-trained model was fine-tuned using a small number of high-confidence samples to predict the retention time and fragment ion intensity of candidate peptides.
[0125] Alternatively, an acceleration strategy or model quantization and half-precision inference techniques can be used to calculate the consistency characteristics between predicted values and experimental observations.
[0126] For open search scenarios, unconventional modifications outside the scope of data coverage are fine-tuned by using additional features as indicators of reduced weighting, automatically lowering their prediction confidence weights.
[0127] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the deep learning-enhanced protein search engine optimization method described above.
[0128] Fourthly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the deep learning-enhanced protein search engine optimization method as described above.
[0129] In addition, combined Figure 1 The deep learning-enhanced protein search engine optimization method described in this application embodiment can be implemented by electronic devices, such as computer devices. Figure 4 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of this application.
[0130] In some embodiments, the computer device may further include a communication interface 83 and a bus 80. For example, Figure 4 As shown, the processor 81, memory 82, and communication interface 83 are connected through bus 80 and complete communication with each other.
[0131] Specifically, the processor 81 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.
[0132] The memory 82 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 81.
[0133] The processor 81 implements any of the deep learning-enhanced protein search engine optimization methods described in the above embodiments by reading and executing computer program instructions stored in the memory 82.
[0134] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0135] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A deep learning-enhanced protein search engine optimization method, characterized in that, The method includes: Peptide spectrum matching coarse scoring steps: For the mass spectrometry data of proteins, hash index is used in peptide spectrum matching to perform accurate matching calculations between theoretical fragment ions and experimental peaks within a set tolerance. A batch processing peptide caching mechanism and a dedicated memory pool management strategy are used in the calculation. Peptide spectrum matching rescoring step: For the mass spectrometry data that has been coarsely scored, a dual-decoy calibration strategy is adopted to divide the decoys into a mutually exclusive visible queue for training and a hidden queue for false discovery rate estimation. A calibration factor is introduced to correct the false discovery rate estimation and output the rescored data. Deep learning steps: Based on the rescoring data, a deep learning model is used to predict the physicochemical properties of peptides. Consistency features are constructed by comparing with experimental observations, and a trust-aware mechanism is adopted to realize deep learning-enhanced rescoring in open scenarios.
2. The deep learning-enhanced protein search engine optimization method according to claim 1, characterized in that, The peptide profiling and scoring step includes: Candidate enumeration and matching steps: The protein database is theoretically digested, and possible modification combinations are enumerated in the specified modification list to generate theoretical peptides. Based on the parent ion mass tolerance, the theoretical peptides are quickly matched and associated with experimental spectra. Batch scoring steps: Whenever a peptide matches a spectrum, the peptide is added to the cached peptide list of the spectrum. When the cached list of a spectrum reaches a predetermined capacity, a batch coarse scoring is performed, and the peptides are added to the candidate peptide list of the spectrum in sequence, while the cached peptide list is cleared; thus realizing that each spectrum corresponds to a batch of peptides. Memory management steps: A memory management mechanism is adopted to initialize a user-mode memory pool for each thread. The size of the memory pool is dynamically expanded at runtime, supporting the creation and destruction of peptide objects.
3. The deep learning-enhanced protein search engine optimization method according to claim 1, characterized in that, The peptide profile matching and rescoring step includes: Dual-decoy segmentation steps: Based on the open and restricted search results, initialize all the decoy peptide-spectrum matching PSMs and divide them into hidden decoys and visible decoys of equal size and mutual exclusion according to certain rules. Rescoring iteration steps: Select non-visible decoys within a specified false discovery rate threshold as positive samples, filter some visible decoys as negative samples, integrate their features, and train the rescoring model. The trained model is used to predict all samples, and the samples are reordered based on the prediction results. The false discovery rate is estimated using hidden decoys. This process is repeated multiple times until convergence or a specified number of times is reached.
4. The deep learning-enhanced protein search engine optimization method according to claim 1, characterized in that, The deep learning steps include: The pre-trained model was fine-tuned using a small number of high-confidence samples to predict the retention time and fragment ion intensity of candidate peptides. Alternatively, an acceleration strategy or model quantization and half-precision inference techniques can be used to calculate the consistency characteristics between predicted values and experimental observations. For open search scenarios, unconventional modifications outside the scope of data coverage are fine-tuned by using additional features as indicators of reduced weighting, automatically lowering their prediction confidence weights.
5. A deep learning-enhanced protein search engine optimization system, employing the deep learning-enhanced protein search engine optimization method as described in any one of claims 1-4, characterized in that, The system includes: Peptide spectrum matching coarse scoring module: For protein mass spectrometry data, the peptide spectrum matching module uses hash index to perform accurate matching calculations between theoretical fragment ions and experimental peaks within a set tolerance, and employs a batch processing peptide caching mechanism and a dedicated memory pool management strategy in the calculation. Peptide spectrum matching rescoring module: For the mass spectrometry data that has been coarsely scored, a dual-decoy calibration strategy is adopted to divide the decoys into a mutually exclusive visible queue for training and a hidden queue for false discovery rate estimation, and a calibration factor is introduced to correct the false discovery rate estimation and output the rescoring data. Deep learning module: Based on the rescoring data, the deep learning model is used to predict the physicochemical properties of peptides, and consistency features are constructed by comparing with experimental observations. A trust-aware mechanism is adopted to realize deep learning-enhanced rescoring in open scenarios.
6. The deep learning-enhanced protein search engine optimization system according to claim 5, characterized in that, The peptide profiling and scoring module includes: Candidate Enumeration and Matching Module: Performs theoretical enzymatic digestion on the protein database, enumerates possible modification combinations in the specified modification list, generates theoretical peptides, and rapidly matches and associates theoretical peptides with experimental spectra based on the parent ion mass tolerance. Batch scoring module: Whenever a peptide matches a spectrum, the peptide is added to the cached peptide list of the spectrum. When the cached list of a spectrum reaches a predetermined capacity, a batch coarse scoring is performed, and the peptides are added to the candidate peptide list of the spectrum in sequence, while the cached peptide list is cleared; thus realizing that each spectrum corresponds to a batch of peptides. Memory management module: Employs a memory management mechanism to initialize a user-mode memory pool for each thread. The size of the memory pool expands dynamically at runtime and supports the creation and destruction of peptide objects.
7. The deep learning-enhanced protein search engine optimization system according to claim 5, characterized in that, The peptide profile matching and rescoring module includes: Dual-decoy segmentation module: Based on the open and restricted search results, initialize all the decoy peptide-spectrum matching PSMs and divide them into hidden decoys and visible decoys of equal size and mutual exclusion according to certain rules; Rescoring Iteration Module: Selects non-visible decoys within a specified false discovery rate threshold as positive samples, filters some visible decoys as negative samples, integrates their features, and trains the rescoring model. The trained model is used to predict all samples, and the samples are reordered based on the prediction results. The false discovery rate is estimated using hidden decoys. This process is repeated multiple times until convergence or a specified number of times is reached.
8. The deep learning-enhanced protein search engine optimization system according to claim 5, characterized in that, The deep learning module includes: The pre-trained model was fine-tuned using a small number of high-confidence samples to predict the retention time and fragment ion intensity of candidate peptides. Alternatively, an acceleration strategy or model quantization and half-precision inference techniques can be used to calculate the consistency characteristics between predicted values and experimental observations. For open search scenarios, unconventional modifications outside the scope of data coverage are fine-tuned by using additional features as indicators of reduced weighting, automatically lowering their prediction confidence weights.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the deep learning-enhanced protein search engine optimization method as described in any one of claims 1-4.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the deep learning-enhanced protein search engine optimization method as described in any one of claims 1-4.